Research Of Video Object Tracking Algorithm Based On Particle Filter Under Complex Environment

Video object tracking, which has broad application prospects in many fields such as navigation guidance, video surveillance, human-computer interaction, medical diagnosis and so on, is a hot research direction in the field of computer vision. The task of video object tracking is to track the moving objects interested in video sequences, and then to obtain the motion parameters of objects, which lay the foundation for further analysis. However, as the real tracking environment is complicated, such as obstruction of objects, deformation of objects themselves, illumination variation, objects motion mutation and so on, it is rather difficult to develop a kind of video object tracking algorithm with high accuracy, good robustness and strong instantaneity. The particle filter algorithm developed in recent years can maintain good robustness and accuracy in nonlinear and non-gaussian system, so more and more scholars apply the algorithm to track video object. Aiming at the problems when using particle filter in complex environment for video object tracking, this paper proposes two improved single maneuvering object tracking algorithms.When the state space dimensions of the object or the number of particles increase, the computational complexity of traditional particle filter method based on multiple features fusion presents exponential growth. By using stratified sampling, this paper presents an adaptive multiple features fusion particle filter algorithm based on stratified sampling. The particles set will be divided into two kinds of particles to describe object state, then each particle only need to compute one feature of the object, so each particle only need to represent a low dimension state, not only retains the multiple features information, but also reduces the computational complexity. In order to highlight the status of the feature which has high credibility, the proposed algorithm allocates more particles to the high credibility feature dynamically according to the its credibility. In addition, the proposed algorithm can adaptively update the fusion coefficients of the features, number and spread scope of particles and feature templates, so the algorithm has a great adaptability to complex tracking scenarios.Aiming at the inherent particle degradation problem of particle filter algorithm and the particles scarcity problem leaded by resampling step, this paper proposes a particle swarm optimization genetic particle filter algorithm based on local multi-zone. The algorithm leads the sampling particles to high likelihood area using the particle swarm optimization, slowing down the particle weight degradation; Then, the algorithm increases the diversity of particles through genetic algorithm instead of the traditional resampling step, avoids algorithm falling into local optimum, enhances the global search ability of algorithm, thus relieves particles scarcity problem. Proposed algorithm merges the latest measurement information into the importance density function, which is made more close to the real posteriori probability distribution of the object. In addition, when the object is obscured, the proposed algorithm randomly selects local area as the particle state model, so that the particle state model can contain block information as little as possible, overcoming the barrier interfering problems effectively; At the same time, describing object by local area can reduce the amount of calculation and improve the instantaneity of algorithm.